As the ARM server market began to emerge in press and powerpoint, it was not hard to separate the hype from reality: it was a lot of hype. Spread by well-meaning advocates trying to change the world and give Intel a run for their money, these myths created unrealistic expectations on whether ARM chips are worthy of server applications, when they will ship, and how hard they will be to use. I applaud the early leaders including APM and AMD for their early efforts on 64-bit products. While they have tried to balance their excitement and the uncertainty of semiconductor development schedules, there are nonetheless a few myths that need clearing up. Here are six common ones: [Read more…]
Well, not really. Not Calxeda, anyway. Just me looking for my favorite audience. Hope you are all still interested in this space, as it is heating up as expected. But now, its not just ARM and a couple little silicon companies. Now, its the big boys, like Intel, AMD, and even IBM. And no, not all are using ARM, but all are innovating more efficient servers for the new hyper scale datacenter.
It has been just over four months since Calxeda collapsed under the combined weight of massive expenses and the competitive challenges it inspired by its vision of low-power ARM-based SOCs for the datacenter. Much as been written about why it failed and what that does or does not mean to the movement it started. But armservers.com started a community of thousands of readers who wan to understand what this David and Goliath battle is all about and what it might mean to the industry. This story has only just begun. Calxeda opened the bottle, and even Intel can’t put the genie back in.
Since I wrote most of the blogs in armservers.com, and since I still have password access to site, I am just going to camp out here, play pirate, board this ship, and start sailing until someone notices and makes me walk the plank to another website. If I get booted off, follow me on twitter @karlfreund, and I will tell you where I set up another soapbox.
For the next few blog postings, here’s what I plan to write about, just to catch everyone up on what’s been going on since our hiatus. Thereafter, I will post musings and perspectives as as new products are rumored, announced, or delayed, and as benchmarks inevitably get published and abused.
- Market Summary
- Current vendor Landscape
- Individual Vendor Assessments
- Intel’s SWOT and Response
But first, I plan to talk about some of the myths and realities of this market. It’s time for me to pay my debts like a Lanister (ignore that reference if you aren’t a fan of Game of Thrones), and come clean from the hype I personally helped create at Calxeda. Not that I was alone, so I will dispel a few myths that my worthy former competitors added to the pile. So my next posting will try to separate the hype from what I think is reality. Good, Bad and Ugly. If I suddenly go quiet, follow me @karlfreund. (or look for me in a dumpster somewhere around Austin! ;-)
Ex-VP Marketing Calxeda, IBM, CRAY, and HP
Well, this is one blog I never thought I would have to write. As you probably know, Calxeda has run out of funding and has closed our offices except a small crew working on a few customer projects. In case you missed it, here is our statement:
Over the last few years, Calxeda has been a driving force in the industry for low power server processors and fabric-based computing. The concept of a fabric of ARM-based servers challenging the industry giants was not on anyone’s radar screen when we started this journey. Now it is a foregone conclusion that the industry will be transformed forever.Now its time to tackle the next challenge. Carrying the load of industry pioneer has exceeded our ability to continue to operate as we had envisioned. We wanted to let you know that Calxeda has begun a restructuring process. During this process, we remain committed to our customer’s success with ECX-2000 projects that are now underway.Calxeda is proud of what we have accomplished, the partners who have collaborated with us, the investors who supported us, and the visionary customers who have encouraged us and inspired us along the way. We will update you as we conclude our restructuring process. In the meantime, we want to thank you personally for your interest and enthusiastic support. Its been an amazing journey.Energy, matter, and innovation are never lost, just reassembled. We look forward to the inevitable application of our ideas.
Calxeda has announced its second generation SoC, the ARM® Cortex™ A15 based EnergyCore™ ECX-2000. This is the industry’s first ARM-based SoC enabled for full OpenStack clouds, Xen and KVM virtualization, and delivers twice the performance of the first generation ARM-based server SoCs. Calxeda will demonstrate the new platform running Ceph object storage and OpenStack at this week’s ARM TechCon conference in Santa Clara, October 29-31. Notably, HP has selected the ECX-2000 for an upcoming Moonshot server in early 2014. Calxeda also added a second 64-bit SoC to its roadmap that is pin-compatible with the ECX-2000, accelerating the availability of production 64-bit Calxeda-based systems in 2014 and protecting customers investments.
While this is big news, there is a far more important story to be told. The new ECX-2000 is just the next step on the journey to a far more efficient datacenter. This journey will fundamentally reshape the datacenter infrastructure into a fleet of compute, storage, networking, and memory resources; the so-called Software-defined Data Center.
Frank Frankovsky, Chairman of the Open Compute Foundation, to join Calxeda Board of Directors: What it means for Calxeda
Calxeda has added the great bearded one to it’s Board of Directors. We couldn’t be more thrilled with Frank’s confidence in Calxeda and the future of our exciting roadmap. Frank will bring an amazing wealth of knowledge and contacts to our business, which can make the difference between being mediocre and being awesome. THINKING you know what customers want is nice, but KNOWING for certain is invaluable to making the right decisions.
Designing a processor is always an art of making trade-offs, and this is especially true when you are designing an SOC, with all the functionality of an entire server on a small slice of silicon no larger than your fingernail. Its not about the processor, its about the entire system, or even a large slice of the entire datacenter. You have to decide whether to dedicate transistors and die area (and therefore cost) to cores, cache, memory controllers, I/O controllers, which I/O features, fabric switches, management engines, fabric software features, management software/firmware, etc, etc, etc. Would your target market customers prefer more cores? Ok, then you better have enough memory controllers to feed them, or they starve (aka, sit around and do nothing). Oh, you want more cores but don’t have room for enough cache? Well, thats a problem. (See a certain new atomic chip with anemic core-to-cache ration for a good example ;-) Again, you are giving the cores time to take a little coffee break while they wait for memory fetches. And that means poor efficiency.
Since we are all about efficiency at scale, these same tradeoffs apply to the all-important features of the interconnect, or fabric. Here, the trade-offs are not the normal domain of a processor designer. These are large system behavioral questions that span networking, compute, and storage. Frank’s datacenter knowledge from his years at Dell DCS and Facebook will give us the inside track to build more than a good chip. We are enabling datacenter infrastructure, or, so-called “software-defined infrastructure” .
All these little decisions will determine the fate of a project, and perhaps a company. There is no margin for error. We are very honored to have Frank join us on this journey and help us become a great company and a vital asset to our customers . Welcome aboard, Frank! Enjoy the ride!
Intel is widely expected to announce a new version of their ATOM SOC for microservers next week. Based on the Silvermont microarchitecture, the Avoton SOC is widely expected to repair their reputation after the disastrous Centerton product that has been largely ignored as a way-too-little-too-late response to ARM.
While we all eagerly await the final specs, and prices, some speculate that this chip will make it harder for ARM-based server SOCs to get traction. I think the opposite is more likely. If this chip is really good, and priced to sell, it means that Intel itself has capitulated to the market demands for a lower power chip designed for real workloads instead of benchmarks. And THAT will validate everything the ARMy of SOC guys have been saying: you don’t always NEED a Xeon behemoth, so why pay for it in terms of power, space, and $$$?? And of course, they wouldn’t do that, at the potential expense of Xeon margins, if they really thought this was less than 10% of the market, and if they didn’t feel threatened by ARM.
So, I’m rooting for Intel for a change. Validate the market, join the party, and may the best RACK win! (a thinly veiled reference to the important of the fabric, which Calxeda fans may appreciate ;-)
Written by Rosemary Francis – Ellexus, Ltd.
The new ARM servers are an economical choice for anyone looking to host massively parallel applications; the technical specifications speak for themselves. But the next question is always about the cost of migrating applications to this new architecture.
Whether moving from x86 or from a PowerPC architecture, you are likely to have to recompile your applications for the new ARM servers. A little preparation early on can save a lot of time in the long run. You can also save time by using Breeze, a tool from Ellexus to help you migrate your applications easily.
Step 1: Work out what your dependencies are
Before you move anything, work out what you need to move. Breeze is a dependency tracing tool that gathers information about your programs as they run. It will show you which programs are used by your application as well as all their files and libraries.
As well as build scripts and source code you’ll want to pre-install all the programs that your build uses. Tracing your application inside Breeze will give you this. You can pull up a list of all the files that have been used as well as all the programs that are called.
Step 2: Recompile your application on the ARM server
If your build uses the default path settings to find the compilers and other programs needed to compile your program, you’ll be able to run your build just fine. Most of the commonly-used programs will be already available and you’ll be able to install them using ‘apt-get install’ or via a similar package manager.
If your scripts have system-specific code to find compilers and libraries then you may need to tweak a few lines to point in the right place. For example, on ARM servers the GNU C libraries will be located in /lib/arm-linux-gnueabihf whereas on a x86 machine they are in /lib64.
The ARM servers are so new that the software partners are bringing out performance-enhancing updates all the time. It is a really good idea to review each of your major dependencies and check that you have the most up-to-date version. Breeze will give you a list of libraries, files and programs used by your build.
What to do if something goes wrong
Occasionally build problems will arise and Breeze can help you to find a solution quickly. Trace your build on the ARM server and compare it with the build on your old machine. There will be some differences that don’t matter, but you should be able to work back from the failure and pick out unexpected differences.
You can also see what is going on inside the build. For example, if you get a “library not found” error from your linker, Breeze can tell you if it is looking in the wrong place or whether it is finding your library and rejecting it.
Step 3: Run your application and profile its performance
Hooray! You are up and running! Now you want to see how your programs perform. IO-bound applications are particularly suited to the ARM servers and you can use Breeze to profile the file IO to gain a holistic overview of performance.
Choose areas of the file system to profile; the tool install directory, project data and temporary storage areas work well. Breeze will then track how much file IO goes to those areas so you can see what the application is doing.
You can pick out peaks of IO when the application loads and when it performs certain tasks. This can be compared with other machines to give you an indication of which areas of the application are faster and which are slower.
The graph above shows the number of bytes written to /tmp and /home in a multi-target build over time. I can use this information to optimise my file system as well as pick out performance bottlenecks. In this way Breeze can make it easy to migrate your applications and get a good insight into the performance implications of the new architecture.
With the allure of a good book on the beach stealing everyone’s attention this summer, you may have missed three important developments in the Linux community to support ARM in the datacenter. The first was the announcement that with the new Fedora 19, Fedora has released ARM and x86 support simultaneously. This was made possible, in part, thanks to a Boston Viridis build-farm installed earlier this year. ARM support is now available in media and installer images for TI OMAP4, nVidia Tegra 2, and Calxeda ECX-1000 (Highbank). This represents a key milestone in providing complete Linux packages for ARM based development, appropriate for customers who roll-their-own OS from open source. (Note that the Ubuntu community already enjoys a fully supported enterprise OS for ARM, thanks to the work of Canonical Ltd, a long-time supporter of the ARM architecture.)
Next on deck for your summer reading pleasure, is the release of Xen 4.3 for ARM V7 and V8 architectures by the Xen Project, enabling hypervisor support for 32- and 64-bit ARM SOCs. Once again, this work was done on a Calxeda-based Boston Viridis system.
Finally, today ARM and Oracle announced the next phase of their collaborative relationship to optimize Javafor ARM-based servers and embedded SOCs, extending their work to 32- and 64-bit optimization:
- Agreement will provide ARM architecture support for key markets e.g. data centers, network infrastructure and embedded computing
- Oracle JVM optimized further for 32-bit products and ported over and optimized for ARMv8 64-bit
- Additional areas for co-operation include improving boot-up performance, power savings and library optimization
Note that Oracle Java SE is a fundamental technology for all of the market areas mentioned above.
Ok, back to the beach…
Live from HostingCon, we’ve posted a video of Brett from Interworx demoing the Interworx control panel and clustering technology on a 24 server Calxeda system. To see the cluster in action for yourself, tweet @InterWorxArm and see what the cluster has to say. We’ll keep the twitter demo running until the end of HostingCon.
Check out the demo video below: